ASI-LIB-036 technical survey

AI Agent Systems: Architectures, Applications, and Evaluation

Bin Xu

AI agent systems survey overview diagram
Figure via ar5iv rendering of arXiv:2601.01743

This survey is useful as scaffolding for the whole library. It organizes agent systems around model cores, memory, world models, planners, tool routers, critics, orchestration patterns, and deployment settings.

Why it matters

Agent evaluation is hard because long-horizon workflows are non-deterministic, tool-dependent, and sensitive to retry budgets, context growth, and environment variation. That is exactly the measurement problem ASI-oriented systems have to solve.

ASI relevance

As agents become the default interface for frontier models, architectural taxonomy becomes operational. It tells us what must improve: memory, planning, tool use, verification, guardrails, and reproducible evals under realistic workloads.